Menu
Book a demo
Machine Vision vs Computer Vision: Which One Does Your Station Need?

Machine Vision vs Computer Vision: Which One Does Your Station Need?

How machine vision and computer vision differ on a plant floor, with a station decision table, supplier questions and a three-station example costed in USD.
Machine Vision vs Computer Vision: Which One Does Your Station Need?

In short: Computer vision is the broad field of getting software to extract information from images. Machine vision is computer vision built into a station so that each unit gets a verdict, such as pass, fail, a measurement or a reject, inside the cycle time. On a plant floor, computer vision also names a second kind of system: cameras with a wider view that count output, time work and show where it waits.

  • Pick machine vision when each unit needs a verdict before it leaves the station. Pick a counting camera when output is made by hand and no machine gives a signal. Pick neither when a PLC or sensor already counts.
  • Price the check you would replace first: human visual checks catch about 80% of defective items on average, and the median US inspector in manufacturing earns $49,270 a year.
  • Example plant: two stations put $143,594 to $214,272 a year in play, a manual seal check and a hand-packing table with no count.

Guide, updated 16 November 2026. Sources are listed at the end.

The two terms get mixed up because the hardware can look alike: a camera on a bracket over a conveyor. What separates them is the decision the camera serves. Below, each kind of station is matched to the system that fits it, starting from the check or the count you would replace, with a worked example that ends in money. Where the job is counting output across a line, our guide to the best computer vision tools for manufacturing OEE compares the options on that side.

What each term means on a plant floor

Computer vision is the field: the methods that turn images or video into information, from reading a road sign to grading fruit. It names the algorithms, not a particular box of hardware.

Machine vision is computer vision built into a machine or a station to make a decision about each unit. Every part of it is engineered around one question. A lens and a light make one feature visible. A trigger fires the camera when the part is in place. Software measures or classifies. Outputs wired to the PLC let a reject gate, a robot or a stop act on the answer before the next unit arrives.

The camera industry writes its standards for this kind of hardware. EMVA 1288, from the European Machine Vision Association, is the standard for measuring and presenting the specifications of machine vision sensors and cameras, with a unified method to measure, compute and present them. Release 4.0 has been in effect since June 2021.

On a plant floor the phrase computer vision has also come to mean a different system: cameras with a wide view of a work area that count output, time work or show where material waits, and report to people and production software rather than to a reject gate. That is the sense used in OEE and light MES.

The methods overlap. Learned models now run inside inspection systems, and counting systems use some of the same tools. The overlap is in the software. The jobs stay different, and so do the price, the proof that the system works, and who owns the result.

Which job is this station?Start from the decision, not the cameraA verdict on each unitbefore it leaves?YesMachine visioncheck and rejectNoDoes a machine givea usable count signal?YesPLC or sensorno cameraNoOutput made by hand,with no signal?YesCounting cameracomputer visionThe baseline you would replace80%average catchof human checks35%good partsrejected (test)$49,270US inspectormedian, mfgSandia National Laboratories, 2015; BLS, May 2025Price the check you replacebefore you price the camera.
Decision flow for one station. Human check figures from Sandia National Laboratories (2015); wage from the BLS (May 2025). See the source.

Explainer presented by Iris, Fabrico's AI-generated presenter: how an add-on camera counts output where a machine has no signal. Book a demo.
Plays from YouTube. Google may set cookies when you press play. Privacy policy.

Two jobs: a verdict on each unit, or a record of the line

Ask what happens to the answer. If it has to change what happens to this unit, in this cycle, it is an inspection job and machine vision is the tool. If it has to tell a person what the line did over a shift, it is a measurement job, and a counting system or a machine signal is the tool.

Inspection (machine vision)Line record (computer vision counting)
Question it answersIs this unit good?How many units left the station, and when did output stop?
TimingInside the cycle, before the next unitOver minutes, shifts and orders
OutputPass or fail, a measurement, a reject or stop signalCounts per minute, gaps in output, totals per shift and order
HardwareIndustrial camera, chosen lens, dedicated light, trigger, I/O to the PLCA camera with a wide view of the station, plus software that turns images into counts
SceneControlled: fixed distance, fixed light, part in a known placeA real work area: people, hands, cases, changing light
Proof that it worksA test set of good and faulty units; misses and false rejects measuredCounts reconciled against a manual tally or pallet labels
Who owns itQuality and process engineeringProduction, CI and the shift leader
What a fault costsBad units shipped, or good units scrappedA wrong OEE, and losses nobody can see

The split also explains where Fabrico sits. Fabrico's AI cameras (an add-on) are counting cameras: they count output at stations where a machine has no usable signal. They are not an inspection system: they do not check units for defects, read codes, recognise stops or propose reasons.

How good is the human check you would replace?

A camera request usually starts from a person: an operator who looks at seals, labels or parts as they pass. Measure that baseline before you price anything, because human visual inspection is less reliable than it feels.

A 2015 study by Sandia National Laboratories, published in the journal Human Factors, sums up decades of research. For most inspection tasks, it reports, missed defects typically run between 20% and 30%, and accuracy of about 70% to 80% holds whatever the product, from piston rings and contact lenses to food.

The study then tested 82 qualified inspectors on 140 parts with eight kinds of defect. They rejected 85% of the defective items, a little above the industry average of 80% that the paper cites. They also rejected 35% of the acceptable parts. Careful inspectors bought their catch rate with scrap.

Two lessons follow for any station.

  1. Measure both errors. Misses send bad units to the customer. False rejects throw good units in the bin. A system that improves one and worsens the other can lose money, and so can a person.
  2. Measure them on a known set. The US Food and Drug Administration's draft guidance on inspecting injectable products for visible particulates (December 2021) asks for each unit to be inspected by trained people, automated equipment or both. It says machine variability is generally easier to control than the variability people bring to repetitive tasks, and that automated inspection machines should be validated to meet or surpass human inspection, using a test set of good and defective units prepared by quality assurance.

The guidance is written for injectables, but the method transfers to a seal or a label check: build a library of real faults, run it past the person and past the system, and compare the two error rates.

Rules, learned models and the images each one needs

Inside machine vision there are two ways to build the check, and they cost effort in different places.

  • Rules set by an engineer. Find an edge, measure a gap, match a pattern, compare with a threshold. Fast, explainable and stable when the light and the part position are fixed. Gauging, presence checks and code reading are usually built this way. Rules break when the product or the scene varies more than they expect.
  • Learned models. A model trained on labelled images learns what a good unit looks like and what a fault looks like. It copes with natural variation, such as wood grain, baked goods or textured plastics, where a rule would need constant tuning.

Learned models move the effort from writing rules to collecting images. A 2025 review from the Indian Institute of Technology Mandi notes that images of normal parts are considerably easier and cheaper to collect than images of faulty ones. That is why one family of methods trains on good units only and treats anything that does not fit as suspect. The review notes that such models can catch new and unforeseen kinds of fault. In practice they can also flag harmless changes, such as film from a new supplier, until they are retrained.

For a manager the questions are practical. Who collects and labels the images? How many examples of each fault exist? Who retrains the model after a product or packaging change, and how does the plant prove the retrained model is at least as good as the last one?

Counting systems use learned models too, but the question they answer is simpler: did a unit leave the station, and when. That simpler question is why counting can work in a busier scene than inspection can.

Which one does this station need? A decision table

Use the table one station at a time. The cost column gives a sourced anchor, not a quote. This guide found no independent, current price survey for vision systems, and supplier prices move with optics, integration and validation. What can be sourced is the cost of the work a system would replace.

OptionGood atCost anchorPick it ifAsk the supplier
A trained person checking by eyeJudgement, new kinds of fault, low volumes, frequent product changes$49,270 median wage a year per inspector in US manufacturing; about $73,647 with benefits (BLS)Volume is low, the fault is rare or hard to define, or the product changes oftenAsk yourself: what share of faults does the person catch on a known test set, and how many good units go in the bin?
Rule-based machine vision (smart camera or vision PC)Presence, position, gauging and code reading at line speedSupplier quote, compared with the manual check it replaces: $49,270 to $73,647 a year per shift positionEach unit needs a verdict, and the light, distance and part position can be fixedIs the camera specified to EMVA 1288? How are recipes switched at changeover? What false reject rate on our products?
Machine vision with learned modelsVariable products and cosmetic faults that rules cannot defineSupplier quote plus the time to collect and label images; same manual baselineFaults vary, and you can collect labelled images of the ones that matterHow many images per fault type? Who retrains after a packaging change, and how is it validated again?
Code reader or verifierReading barcodes and 2D codes; grading print qualitySupplier quote; the cost of a no-read is a stopped or relabelled caseCodes must be read on each unit, or a customer requires a print gradeIs it a read rate only, or grading to ISO/IEC 15416 or 15415 with a verifier?
Machine signal (PLC or sensor)Counting cycles and units, run time and stop timeNo camera: it reads a signal the machine already hasThe machine already gives a cycle or count signalWhich signal marks one good unit, and are rejects counted separately?
Computer vision counting cameraCounting output at manual stations, gaps in output, timingSupplier quote, compared with the paid hours a count exposes: $25.03 to $37.21 per worker hour (BLS)Output is made by hand and no machine gives a signalWhat must be in view? How are counts reconciled with the shift total? Are people in view, and where do images go?

Then read off the verdict for the station:

  • Choose machine vision if each unit needs a verdict before it leaves the station and you can fix the light, the distance and the part position.
  • Choose learned models inside machine vision if the faults vary too much for rules and you can collect labelled images of them.
  • Choose a counting camera if output is made by hand, no machine gives a usable signal, and the question is how much was made and when work stopped.
  • Choose the machine signal if the machine already counts. A camera adds nothing to a count the PLC already has.
  • Keep a trained person if the fault is rare, hard to define or changes with each product, and measure that person on a test set like any other method.

Where both jobs exist on one line, they are separate purchases. An inspection system will not tell you why the line made 12% less than plan, and a counting camera will not keep a bad seal off a pallet. For where inspection sits in a quality plan, see inline vs offline inspection and AQL vs 100% inspection.

Worked example: three stations on one snack packing line

Example plant. Every figure in this section is example data except the unit costs, which come from the US Bureau of Labor Statistics. The plant packs snack bags into cases on one line, on two 8-hour shifts with 7.5 planned hours each, 250 days a year. The production manager has a request for cameras on three stations.

Station A: the seal and label check at the end of the line

Today one checker per shift watches the bags pass and pulls any with a bad seal or a crooked label. Each bag needs a verdict before it is cased, so this is an inspection job: machine vision, or keep the person.

First, the cost of the manual check. The BLS Occupational Outlook Handbook puts the median wage of inspectors, testers, sorters, samplers and weighers in manufacturing at $49,270 a year (May 2025). Wages are 66.9% of total compensation in manufacturing, according to the BLS employer costs table for June 2026, so one position costs about 49,270 / 0.669 = $73,647 with benefits.

  • Wages only: 2 positions x $49,270 = $98,540 a year.
  • With benefits: 2 positions x $73,647 = $147,294 a year.

Then the quality side. The line packs 3,000 bags an hour, so 2 x 7.5 x 3,000 = 45,000 bags a day. If 0.4% have a seal or label fault, that is 180 faulty bags a day. At the 20% to 30% miss rate the Sandia paper calls typical, 36 to 54 of them reach a case every day with the checker in place.

The decision: ask suppliers for an all-in quote for a vision station with a reject. It should cover the camera, the lighting, the reject gate, integration with the line, a recipe for each bag size and validation on a test set of real faulty bags. Over two years the manual check costs 2 x $98,540 = $197,080 to 2 x $147,294 = $294,588, which is the ceiling for that quote on a simple two-year payback. The saving is only real if the checkers' hours move to work the plant needs, such as AQL samples and changeover checks.

Station B: the barcode and date on each case label

A labeller prints a GS1 barcode with the lot and best-before date on each case. Each case needs the right label, so this is also a per-unit job, but a narrower one: a code reader at the labeller, checked against the day's lot.

Reading is not grading. GS1 Sweden's guidance on barcode quality says GS1 standards require barcodes to reach ANSI grade C, which corresponds to ISO/IEC 15416 grade 2, when the item reaches its best-before date. A verifier measures that grade. A reader on the line only proves the code could be read today. Before buying a reader, check the basics GS1 Sweden lists: printer heads cleaned at regular intervals, label material that resists wear and moisture, and clear space on both sides of the code. The money on this station is a quality and customer question, so it stays out of the sums. For the trade-off between code checks and counting, see label and code vision vs an OEE camera.

Station C: the hand-packing table

Three packers move bags from a buffer table into cases and stack the cases on a pallet by hand. Nothing on the table has a PLC, and the shift's output is a number written on a sheet at the end of the shift. No bag needs a verdict here. The question is how much was packed and when packing stopped, so this is a counting job. A photo-eye would work if finished cases left one at a time on a conveyor. Here they go straight to a pallet, so the choice is a counting camera.

After a month of counts, the table shows on average 2.4 hours a day, across both shifts, with no case packed: waiting for bags after a jam upstream, label roll changes, pallet swaps. That is 3 packers x 2.4 hours x 250 days = 1,800 paid hours a year with no output. The same BLS table gives the hourly cost of production occupations.

  • At the wage of $25.03 an hour: 1,800 x $25.03 = $45,054 a year.
  • At total compensation of $37.21 an hour: 1,800 x $37.21 = $66,978 a year.

The count also gives the table an OEE for the first time. Planned time is 15 hours a day. With 2.4 hours of no output, availability is (15 minus 2.4) / 15 = 84%. At a standard rate of 360 cases an hour, the 12.6 run hours should give 4,536 cases; the count shows 3,950, a performance of 87.1%. Quality comes from Station A's records: 99.5% of cases need no rework. OEE = 0.84 x 0.871 x 0.995 = 72.8%.

A count does not recover the money by itself. It shows where the 2.4 hours go, so the upstream jam and the label changes can be fixed in order of size. How counts become availability and performance is covered in OEE with computer vision.

What the three decisions add up to

Example plant: money in play a yearExample data; unit costs from BLSWagesWages plus benefitsStation A: manual seal and label check$98,540$147,294Station C: paid hours with no output$45,054$66,978Both stations: $143,594 to $214,2722 checkers on 2 shifts; 1,800 idle paid hoursStation B, the case label code, is a quality call
Example data. Unit costs from the BLS Occupational Outlook Handbook (May 2025 wages) and Employer Costs for Employee Compensation (June 2026).

Station A puts $98,540 to $147,294 a year on the table and Station C $45,054 to $66,978. Together that is 98,540 + 45,054 = $143,594 to 147,294 + 66,978 = $214,272 a year in play. Station B is a quality and customer decision with no labour sum. Two different systems, bought for two different reasons, from two different budgets.

Book a demo and see OEE and losses per line and shift.

Questions to put to a supplier before you sign

For an inspection system

  1. Which faults, at which size, at what line speed? Get it in writing for each product, not as a trial on the supplier's best sample.
  2. What are the misses and the false rejects on our test set? Supply real faulty units and good ones, as the FDA guidance describes for automated inspection, and ask for both numbers.
  3. Is the camera specified to EMVA 1288? A shared measurement method lets you compare datasheets from different makers like for like.
  4. Does the camera support GenICam? GenICam, another EMVA standard, gives a generic programming interface for cameras whatever the interface technology, such as GigE Vision, USB3 Vision or CoaXPress, which keeps you less tied to one software stack.
  5. How are recipes switched at changeover, and who can change a limit? A check that runs the wrong product's limits passes bad units with a green light.
  6. Who retunes or retrains after a packaging change, and how is the change validated?
  7. Where do rejects go, and are they counted? Rejected units are quality loss in OEE. Unless that count reaches production records, the loss stays invisible.

For a counting system

  1. What must be in view, and what breaks the count? A person standing in the way, cases stacked high, glare from a window.
  2. How are counts reconciled? Ask for a week in which the count is checked against pallet labels or a manual tally.
  3. What does it give besides the count? Gaps in output, the start and end of an order, rate against standard.
  4. Are people in view? If so, data protection law applies. The European Data Protection Board's guidelines 3/2019 on video devices, adopted on 29 January 2020, say an employee at work is in most cases not expecting to be monitored by the employer, that employers should in most cases not rely on consent, and that national law or works agreements can set specific rules. The same guidelines class simple counting algorithms as less intrusive than complex biometric technologies. Ask where images are processed, who can see them and whether anything is kept, and talk to employee representatives before the first camera goes up.

For mounting heights and light at either kind of station, see camera placement and lighting on a production line.

Mistakes that make the wrong system look right

  • Buying a camera for a count the PLC already has. If a filler or a case packer already gives a cycle signal, read it.
  • Expecting a counting camera to judge quality. Counting and inspection need different optics, lighting, timing and proof. A system validated for one job is not validated for the other.
  • Judging an inspection system on its catch rate alone. The Sandia inspectors caught 85% of faulty parts and rejected 35% of good ones. A supplier who quotes only the first number is quoting half the result.
  • Testing on the supplier's samples. Faults collected from your production over several weeks are the only fair test.
  • Forgetting changeovers. Each product or pack size needs its own recipe, and the recipe has to change with the product.
  • Leaving the reject bin out of OEE. Good units thrown away by false rejects are output the line made and cannot sell.

What to do this week and this quarter

This week

  1. Plant or production manager: list each station where someone has asked for a camera, and write down the decision it serves: a verdict on each unit, or a record of output.
  2. Quality manager: take 50 good and 20 faulty units from one manual check and run them past two people, blind. Record misses and false rejects. That is the baseline any system has to beat.
  3. CI manager: for each manual station, find out where today's output number comes from and when it is written down.

This quarter

  1. Price the manual checks with the BLS figures above or your own payroll, so every quote is compared with a number.
  2. Build a fault library for each station that needs inspection, held by quality and added to whenever a new fault turns up.
  3. Run one station of each kind with a supplier, judged against the questions above, before extending to the rest of the line.
  4. Put counts and signals in one view. Counts from manual stations and signals from machines belong in the same OEE and losses by line and shift, so the reject bin, the idle hours and the stops sit on one chart. OEE software builds that view from machine signals, and a light MES ties the hours to orders and products.

How Fabrico fits

Fabrico is a cloud OEE and light MES platform: OEE and losses by machine, line and shift, from the PLC or a Fabrico IoT device, with reasons operators add on a tablet or phone.

On manual stations, AI cameras (an add-on) count output.

Book a demo.

What people ask next

Is machine vision the same as computer vision?

No. Computer vision is the broad field of getting software to extract information from images. Machine vision is computer vision built into a station so that each unit gets a verdict, such as pass, fail, a measurement or a reject, inside the cycle time. On a plant floor, computer vision also names counting and tracking systems that report what a line did over a shift.

Which do I need for a manual packing station?

Start from the decision. If each pack needs a check before it leaves, that is machine vision, or a trained person. If the question is how much was packed and when work stopped, and no machine gives a signal, that is a counting camera. If a conveyor or machine already gives a count signal, read that signal instead.

Can one camera both count output and check quality?

Plan for two systems. Inspection needs a controlled scene, a trigger, a reject and validation on a test set of faulty units. Counting needs a wide view of a real work area and counts reconciled against a tally. A system validated for one job is not validated for the other.

How accurate is a human visual inspector?

Less accurate than it feels. A 2015 Sandia National Laboratories study cites missed defects of 20% to 30% as typical and accuracy of about 70% to 80% across products. In its own test, 82 qualified inspectors caught 85% of defective parts and rejected 35% of good ones.

How much does a machine vision system cost?

It depends on the optics, lighting, reject, integration, recipes and validation, so ask for an all-in quote. Compare it with the work it replaces: in US manufacturing the median inspector wage was $49,270 a year in May 2025, about $73,647 with benefits, for each shift position (BLS).

How do I compare two machine vision suppliers?

Give both the same test set of good and faulty units from your line, and compare misses and false rejects, not the catch rate alone. Ask for camera data measured to EMVA 1288, GenICam support, and how recipes are switched and validated again after a product or packaging change.

Sources

Sources accessed on 16 November 2026 unless marked as accessed on another date. Dates are the publication dates shown on each page.

Track OEE and downtime in real time
See where lines lose time
Book a live demo.
See where lines lose time
See where lines lose time

Real Fabrico tiles, sample data.

By clicking the Accept button, you are giving your consent to the use of cookies when accessing this website and utilizing our services. To learn more about how cookies are used and managed, please refer to our Privacy Policy and Cookies Declaration